Machine learning-driven multimodal optimization of selenium biotransformation and flavor profiling in fermented apple–Yacon functional beverages

Abstract

This study developed an artificial intelligence (AI)-driven framework to optimize selenium-enriched Yacon-apple juice fermentation. By integrating response surface methodology (RSM) and extreme gradient boosting (XGBoost) modeling, key parameters (34.8 °C, 1:2.2 apple:yacon ratio, 0.65 g/L enzyme) were identified, resulting in 89.78 % selenium conversion and high bioactive yields (149.42 mg/100 mL polysaccharides; 1.250 mg/mL flavonoids). XGBoost demonstrated superior predictive accuracy (R² = 0.953) over traditional RSM, revealing temperature thresholds (34–35 °C) critical for Lactiplantibacillus plantarum YKX (L. plantarum YKX) activity. Headspace-gas chromatography–ion mobility spectrometry (HS–GC–IMS) analysis revealed fermentation-driven flavor evolution: 442 % ester accumulation (ethyl acetate) at 4 days correlated with sensory improvement (r = 0.91), whereas the content of aldehydes decreased by 23 %. Multimodal machine learning linked polysaccharide metabolism to flavor enhancement (R² = 0.927) and identified a critical 12–24 h selenium conversion window (1.52 %/h rate). This work bridges empirical optimization with explainable AI, providing actionable guidelines (temperature control ±0.5 °C, enzyme synergy) for scaling functional foods. Limitations in dataset size highlight the need for sensor-augmented monitoring. This approach advances precision fermentation technologies to balance nutrient bioavailability, flavor complexity, and bioactive retention in selenium-enriched beverages.